BoostedTreesCalculateBestGainsPerFeature

public final class BoostedTreesCalculateBestGainsPerFeature

Calculates gains for each feature and returns the best possible split information for the feature.

The split information is the best threshold (bucket id), gains and left/right node contributions per node for each feature.

It is possible that not all nodes can be split on each feature. Hence, the list of possible nodes can differ between the features. Therefore, we return `node_ids_list` for each feature, containing the list of nodes that this feature can be used to split.

In this manner, the output is the best split per features and per node, so that it needs to be combined later to produce the best split for each node (among all possible features).

The length of output lists are all of the same length, `num_features`. The output shapes are compatible in a way that the first dimension of all tensors of all lists are the same and equal to the number of possible split nodes for each feature.

Public Methods

static BoostedTreesCalculateBestGainsPerFeature
create ( Scope scope, Operand <Integer> nodeIdRange, Iterable< Operand <Float>> statsSummaryList, Operand <Float> l1, Operand <Float> l2, Operand <Float> treeComplexity, Operand <Float> minNodeWeight, Long maxSplits)
Factory method to create a class wrapping a new BoostedTreesCalculateBestGainsPerFeature operation.
List< Output <Float>>
gainsList ()
An output list of Rank 1 tensors indicating the best gains for each feature to split for certain nodes.
List< Output <Float>>
leftNodeContribsList ()
A list of Rank 2 tensors indicating the contribution of the left nodes when branching from parent nodes (given by the tensor element in the output node_ids_list) to the left direction by the given threshold for each feature.
List< Output <Integer>>
nodeIdsList ()
An output list of Rank 1 tensors indicating possible split node ids for each feature.
List< Output <Float>>
rightNodeContribsList ()
A list of Rank 2 tensors, with the same shape/conditions as left_node_contribs_list, but just that the value is for the right node.
List< Output <Integer>>
thresholdsList ()
An output list of Rank 1 tensors indicating the bucket id to compare with (as a threshold) for split in each node.

Inherited Methods

Public Methods

public static BoostedTreesCalculateBestGainsPerFeature create ( Scope scope, Operand <Integer> nodeIdRange, Iterable< Operand <Float>> statsSummaryList, Operand <Float> l1, Operand <Float> l2, Operand <Float> treeComplexity, Operand <Float> minNodeWeight, Long maxSplits)

Factory method to create a class wrapping a new BoostedTreesCalculateBestGainsPerFeature operation.

Parameters
scope current scope
nodeIdRange A Rank 1 tensor (shape=[2]) to specify the range [first, last) of node ids to process within `stats_summary_list`. The nodes are iterated between the two nodes specified by the tensor, as like `for node_id in range(node_id_range[0], node_id_range[1])` (Note that the last index node_id_range[1] is exclusive).
statsSummaryList A list of Rank 3 tensor (#shape=[max_splits, bucket, 2]) for accumulated stats summary (gradient/hessian) per node per buckets for each feature. The first dimension of the tensor is the maximum number of splits, and thus not all elements of it will be used, but only the indexes specified by node_ids will be used.
l1 l1 regularization factor on leaf weights, per instance based.
l2 l2 regularization factor on leaf weights, per instance based.
treeComplexity adjustment to the gain, per leaf based.
minNodeWeight minimum avg of hessians in a node before required for the node to be considered for splitting.
maxSplits the number of nodes that can be split in the whole tree. Used as a dimension of output tensors.
Returns
  • a new instance of BoostedTreesCalculateBestGainsPerFeature

public List< Output <Float>> gainsList ()

An output list of Rank 1 tensors indicating the best gains for each feature to split for certain nodes. See above for details like shapes and sizes.

public List< Output <Float>> leftNodeContribsList ()

A list of Rank 2 tensors indicating the contribution of the left nodes when branching from parent nodes (given by the tensor element in the output node_ids_list) to the left direction by the given threshold for each feature. This value will be used to make the left node value by adding to the parent node value. Second dimension size is 1 for 1-dimensional logits, but would be larger for multi-class problems. See above for details like shapes and sizes.

public List< Output <Integer>> nodeIdsList ()

An output list of Rank 1 tensors indicating possible split node ids for each feature. The length of the list is num_features, but each tensor has different size as each feature provides different possible nodes. See above for details like shapes and sizes.

public List< Output <Float>> rightNodeContribsList ()

A list of Rank 2 tensors, with the same shape/conditions as left_node_contribs_list, but just that the value is for the right node.

public List< Output <Integer>> thresholdsList ()

An output list of Rank 1 tensors indicating the bucket id to compare with (as a threshold) for split in each node. See above for details like shapes and sizes.